Build a useful investigation brief
Reconstruct the frame
Collect population window, task and output unit, filters, metadata, selection probabilities or reasons, reviewer assignment, context, labels, and nonresponse. Mark excluded or unavailable groups and preserve the original sample unchanged.
Compare sample to population
Inspect task mix, risk, time, model version, confidence, disagreement, and outcome distributions. Identify overrepresentation, missing strata, convenience selection, or reviewer attrition. Separate sampling bias from inconsistent human labels within the sample.
Test a correction
Draw a matched random, stratified, or missing-group sample. Compare labels and conclusion. State whether the original result holds, changes, or remains exploratory because the population frame cannot be recovered with confidence.
What to carry forward
The investigation brief should link population frame, selection, reviewed cases, missing groups, labels, correction sample, and conclusion. End with a supported inference or bounded sampling uncertainty. Do not generalize a targeted sample to the full population. Record the selection reason, missing group, and reviewer owner before stating prevalence.
Technical background: Google DeepMind evaluation research.
Keep the decision with the work.
Use a Work Item in Aglet to record the problem, the evidence you have, and the next decision. Add an owner and priority, then keep updates in the discussion so the next person can follow the reasoning.
Create an account See the product workflow